摘要
针对目前三维人脸识别在时间成本和识别精度的平衡问题上,提出了一种基于残差网络的三维人脸识别方法.该方法首先定义了一个二维平均人脸特征点和一个三维平均人脸特征点,将三维点云向三维平均人脸特征点对齐后,做统一的透视投影得到深度图像,再经过人脸区域裁剪得到用于训练的数据,最后使用27层的残差网络训练分类模型,从而实现三维人脸识别.由于提前设计平均人脸特征点,故大幅缩短了数据预处理时间,在FRGCv2.0数据集上进行测试取得了很好的效果:中性对中性实验、对全部实验、对非中性实验,识别率分别为98.8%、98.5%、98.5%,且总耗时仅为0.5秒.
In order to balance the time cost and accuracy of 3d face recognition, a fast 3d face recognition method based on deep learning is proposed. This method predefined a 2D mean face key points matrix and a 3D mean face key points matrix to make 3d point cloud align with the 3D mean face, after unified perspective projection, depth image is thus obtained;through local clipping of face region the data use for training is acquired;Finally the 27-layer Residual Network training classifying model is used thus to achieve rapid 3d face recognition. Since the mean face feature points were designed in advance, the pre-processing time of this method was greatly shortened, and the experiment on FRGC v2.0 data set produced great results with a lot higher accuracy and far less recognition time.
作者
张笑楠
张自友
ZHANG Xiaonan;ZHANG Ziyou(National Key Laboratory of Fundamental Science on Synthetic Vision,Sichuan University,Chengdu,Sichuan 610064,China;Wisesoft Co.,Ltd.,Chengdu,Sichuan 610064,China;School of Physics and Electrical Engineering,Leshan Normal University,Leshan,Sichuan 614000,China)
出处
《内江师范学院学报》
2019年第6期61-67,共7页
Journal of Neijiang Normal University
基金
国家重大仪器开发专项《高速高精度结构光三维测量仪器开发与应用》(2013YQ490879)